Abstract: FR-PO992
Deep Learning-Based Segmentation Enables an Efficient Assessment of Glomerulosclerosis That Is Predictive of Progressive Kidney Disease
Session Information
- Pathology and Lab Medicine: Basic
November 08, 2019 | Location: Exhibit Hall, Walter E. Washington Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pathology and Lab Medicine
- 1602 Pathology and Lab Medicine: Clinical
Authors
- Palmer, Matthew, University of Pennsylvania, Philadelphia, Pennsylvania, United States
- Jayapandian, Catherine P., Case Western Reserve University, Cleveland, Ohio, United States
- Zee, Jarcy, Arbor Research Collaborative for Health, Ann Arbor, Michigan, United States
- Sekulic, Miroslav, UH Cleveland Medical Center, Cleveland, Ohio, United States
- Cassol, Clarissa Araujo, The Ohio State University, Columbus, Ohio, United States
- Rule, Andrew D., Mayo Clinic, Rochester, Minnesota, United States
- Janowczyk, Andrew, Case Western Reserve University, Cleveland, Ohio, United States
- Chen, Yijiang, Case Western Reserve University, Cleveland, Ohio, United States
- Sedor, John R., Cleveland Clinic, Cleveland, Ohio, United States
- Hewitt, Stephen M., National Cancer Institute, Bethesda, Maryland, United States
- Madabhushi, Anant, Case Western Reserve University, Cleveland, Ohio, United States
- Barisoni, Laura, Duke University, Durham, North Carolina, United States
Background
The percentage of globally sclerotic glomeruli (GSG) adjusted by age is a clinically relevant parameter that has been shown to be associated with outcome across diseases. While annotation of glomeruli on whole slide images (WSI) using the NEPTUNE Digital Pathology Protocol has improved overall accuracy, manual counting remains time consuming. The aim of this study is to develop deep learning (DL) networks for automated annotation of GSG and test whether the DL-generated % GSG associates with clinical outcome.
Methods
126 WSI (PAS) from 107 minimal change and 19 FSGS patients from the NEPTUNE dataset were used to train and test a DL network to identify normal glomeruli and GSG. The %GSG was calculated on 1 level and compared with %GSG visually assessed on the same level by 1 of 4 pathologists. Outcome data (ESRD or 40% eGFR decline) were available in 125 cases. Cases were divided into 3 groups: no GSG (95), GSG appropriate for age (14), and GSG excessive for age (16). Hazard ratios for clinical outcomes were compared across the 3 groups.
Results
The DL classifier’s sensitivity as compared with visual assessment for detecting non-GSG was 0.85 and for GSG was 0.75. Compared with no GSG, GSG normal for age was associated with 1.06 (0.13-8.62) times the hazards of composite progression outcome and GSG abnormal for age was associated with 4.43 (1.14-17.26) times the hazards of composite progression outcome.
Conclusion
Our DL classifier is able to detect normal glomeruli and GSG with high sensitivity. The DL-generated %GSG correlates with clinical outcome in a manner similar to previously reported manual assessments. This work represents a foundation towards enlisting robust machine learning classifiers for evaluating clinically relevant pathologic parameters.
Funding
- Private Foundation Support